Reaction-Diffusion Equation-based Model For Image Super-Resolution

The task of image super-resolution reconsturtion is a hot topic in the field of image processing.In recent years,with the introduction of deep learning algorithm,the quality of superresolution image has stepped into a new level.In order to obtain super-resolution results with high Peak Signal to Noise Ratio and Structural Similarity,the existing algorithms focus on the structural design of the deep model,and the gradually complex deep model constantly refreshes the quality of the super-resolution image.However,these models lack the way of local pattern generation in super-resolution,and they rely on the convolution to complete the task of pixel reconstruction.This makes the depth of the model too high,thus leading to difficulty of the application of the super-resolution model.In order to solve this problem,this paper utilizes the pattern generation mechanism of reaction-diffusion equation to reduce the difficulty of super-resolution task,which greatly reduces the number of model parameters and model depthWe study the mechanism of pattern generation of reaction-diffusion equation,then the reaction-diffusion equation is introduced to the image super-resolution area to form the reactiondiffusion module.The newly designed module guide the image super-resolution reconstruction with the use of the local pattern generation ability of reaction-diffusion equation,thus effectively reduces the diffculty of the super-resolution problem.In this paper,a multi-stage cascaded reaction-diffusion framework is designed,which divides the super image reconstruction into several stages,and makes the different depth of model complete different stages.This framework can spread the pressure of image reconstruction to different levels of the model,and can effectively utilize the depth of neural network.Based on the idea of parameter prediction,a learnable reaction-diffusion module is designed.The neural network is used to estimate the diffusion function and reaction function of the reaction-diffusion equation,which improves the ability of pattern generation of the reaction-diffusion module.We propose a light-weight super-resoution network based on the learnable reaction-diffusion equation.By decomposing the super-resolution problem with the reaction-diffusion module,we can reduce the difficulty of the super-resolution task and help the model use fewer parameters to achieve better super-resolution results.Experiments are carried out on the commonly used super-resolution data set.The experimental results verify the feasibility of embedding of the reaction diffusion equation into deep learning.The model proposed in this paper only needs less parameters and shallower network to achieve the same super-resolution performance,which proves the superiority of embedding the reaction-diffusion equation into the depth model.